Split the review in two. The mechanical half is checkable by machine and should be; the judgment half is the reviewer's, and it is also where a junior learns anything.
Key takeaways
Use AI on the half of the review that is a checklist and keep the other half for yourself. Almost every review of junior work contains two different activities that get done in one pass and therefore compete: confirming that the file is complete, internally consistent, arithmetically sound and properly cross-referenced; and forming a view on whether the work was the right work and the conclusion holds. The first is mechanical and a machine is better at it than a tired manager at seven in the evening. The second is the job.
Running them in that order, rather than together, is the whole technique. A reviewer who opens a file that has already passed the mechanical pass starts on the substantive question instead of arriving at it after forty minutes of arithmetic.
What must not go on it is anything that requires a view: whether a judgment was reasonable, whether the scope was right, whether a client explanation is plausible. A checklist that includes those items has quietly moved the judgment to the tool, and the reviewer will start trusting the green ticks.
The AASB's summary of its Canadian Technology Quality Management roundtable names the specific thing firms find hard: how to demonstrate effective direction, supervision and review, professional judgment and professional skepticism when AI is involved. That is the concern in one sentence. Direction, supervision and review is a chain of human attention, and inserting an automated pass into it either strengthens the chain by removing noise or weakens it by manufacturing false comfort. Which one you get depends entirely on whether the checklist stops at the mechanical.
The same summary records the safeguards firms describe: human involvement at critical points, validation of inputs and outputs for engagement-level tools, limits on how AI tools may be used in engagement files, and mandatory AI training before staff can use the tools with ongoing training as tools evolve. For a review workflow the training point is the one to take seriously — a junior who does not understand what the mechanical pass does not check will assume it checked everything.
Here is the part that does not appear in a business case. A junior learns by producing something imperfect and having a senior show them, in the file, why it was imperfect. If the mechanical errors are caught and corrected before the reviewer ever sees them, the junior never receives that feedback, and the path from junior to reviewer gets longer.
The AASB roundtable summary is explicit that this cuts the other way from the productivity story: among its key takeaways is that human oversight remains essential and that skills such as judgment, skepticism, accountability, coaching and mentoring are increasingly important for responsible AI use. It also records ‘supervision and coaching’ as a future use case — smart agents providing proactive prompts and guidance to staff — and a training concern about how to develop the critical thinking skills of junior auditors.
The workable answer is to route the mechanical exceptions to the preparer, not the reviewer. The junior sees their own error list, fixes it, and resubmits. The reviewer sees the file after correction and also sees the exception count over time, which is a far better development conversation than a marked-up schedule.
A firm in Ontario with four managers and eleven staff has a review bottleneck: managers spend the first half of every review on casting and cross-references and the second half rushing the judgment. Files come back to preparers an average of three times.
The firm introduces a pre-review pass that runs the six mechanical checks above and returns the exceptions to the preparer, with a rule that a file cannot be submitted for manager review with open mechanical exceptions. Managers are told what the pass does not check, and the checklist is deliberately kept free of judgment items.
Three things are countable before and after: mechanical exceptions per file at first submission, the number of review cycles per file, and the proportion of manager review comments that concern judgment rather than mechanics. The third is the one that tells you whether the change worked.
Responsibility does not delegate. In accounting that is the architecture the AASB describes — CSQM 1 paragraph 32(f) makes appropriate technological resources a firm-level quality objective, and CAS 220 places engagement-level responsibility on the engagement partner and team, with firm-certified tools treated as engagement resources. In law the same principle is written directly into rule 6.1-1 of the Law Society of Ontario's Rules of Professional Conduct, which requires a lawyer to assume complete professional responsibility for their practice and to directly supervise non-lawyers to whom tasks are assigned, with commentary requiring review of that work at frequent intervals. A tool is not a non-lawyer, but it is even less capable of bearing responsibility, so the supervision standard is the floor rather than the ceiling.
If a claim ever follows, the firm is generally in the frame alongside the individual — see Treadstone Law on suing a professional's firm rather than the individual. A documented review process is what a firm has instead of an argument.
Can AI do the whole review on a low-risk file?
No. It can complete the mechanical pass on any file. The judgment pass is what the sign-off is about, and low risk is itself a judgment.
Should juniors see the AI exceptions?
Yes, and preferably before the reviewer does. That is how the coaching loop survives.
What is the first check to automate?
Sourcing. Flagging assertions with nothing attached costs nothing to run and finds the errors that matter most.
A 30-minute call is enough to tell you whether AI pays for itself here.
